Harnessing the Power of Nondeterminism for Software Engineering
summary
Claims that AI yields a 10x developer boost meet hard examples: token sampling and floating-point nonassociativity explain nondeterminism, and Claude and ChatGPT are the actual tools tried. A tolerance-for-ambiguity matrix (risk vs ambiguity) decides use cases; outcomes: lint-rule discovery stays exploratory with AI, violations get fixed into durable source code, and an iOS/Android GPU animation prototype ships in weeks not months but leaves no in-house metal/vulkan expertise.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:02Non-determinism in software
- 3:51AI non-determinism challenges
- 5:40AI tools limitations
- 8:32LLMs vs compilers
- 11:00Tolerance framework for AI
- 13:13Tolerance axes explained
- 14:59Four quadrants of tolerance
- 17:26Balancing AI takeaways
- 17:47Decomposing problems for AI
- 20:56Durable AI improvements
- 23:21Testable AI outputs
- 25:47AI 3D animation example
- 34:18Summary and Q&A
related talks
01mIdan Gazit, GitHub, ETLS Vegas (29)
37mDeterministic modernization in an age of AI
53mThe Human Side of AI: Unlocking Opportunity in the New Bottlenecks of Software Development
24mThe Impact of AI on the future of software engineering
36mEngineering Intent: Solving the AI Productivity Paradox
26mAI-augmented SDLC: Rethinking how we deliver software
45mThe AI Software Architect: Revolution or Illusion?
44mThe Future of Engineering Teams in an AI-Native World